LLMs: a bleak future ahead?
Why I'm excited about most applications of AI, but have misgivings about the monetization of LLMs.
On most days, I detest punditry. That said, the emergence of large language models (LLMs) is perhaps the most interesting development in my entire career in tech — so it’s simply too consequential to ignore.
I’m not inclined to speculate about the advent of artificial general intelligence (AGI): a hypothetical set of algorithms that match or surpass humans at any given task. I’m unconvinced that we’re nearing that threshold; in an earlier post, I remarked the following:
“The technology feels magical and disruptive, but we felt the same way about the first chatbot — ELIZA — and about all the Prolog-based expert systems that came on its heels. This isn’t to say that ChatGPT is a dud; it’s just that the shortcomings of magical technologies take some time to snap into view.”
The bottom line is that LLMs are designed to function as text predictors. It remains to be seen if their humanlike behavior is merely a parlor trick, a straightforward if unexpected consequence of the vastness of the internet; or if they exhibit some yet-unknown emergent property that sets us on a path toward true AGI.
Instead of taking sides in that debate, I’d like to make a simpler prediction about LLMs as they operate today. I suspect that barring urgent intervention, within two decades, most of interactions on the internet will be fake.
It might seem like an oddly specific claim, but there are powerful incentives to use LLMs to generate inauthentic content on an unprecedented scale — and there are no technical defenses in sight. Further, one of the most plausible beneficial uses of LLMs might have the side effect of discouraging the creation of new organic content on the internet.
To explain my point of view, let me start with the parable of customer support. The need to devote resources to this task is seen by many large companies as a barrier to growth, so over the past three decades, we’ve seen the proliferation of maze-like phone IVRs, chatbots, and support wizards on the web. The first versions of such systems offered a simple way to reach a human. Today, most designs are more adversarial and escape-proof: for example, in an IVR, pressing “0” to speak to a human might just get you chided by a robot that refuses to cooperate. (For a taste, call UPS at 800-742-5877.)
Even with these countermeasures in place, human support continues to be a drag on revenue. Many businesses try to reduce costs by outsourcing the job to third-party call centers where low-wage workers with little agency must follow a rigid script. On the backend, the decision to offer a refund or make other accommodations is seldom predicated on what the customer has to say.
Herein lies the allure of LLMs: platforms such as ChatGPT are already capable of replacing these vestiges of customer support with something that feels more “authentic” to most, at fraction the cost. I have no doubts that LLMs will be employed for this purpose soon, and that most of the time, we won’t even know.
The approach will expand far beyond product support; a recent Twitter thread from a Princeton-educated SF Bay Area entrepreneur offered a glimpse of what’s to come:
The author boasted about the improved metrics and higher satisfaction scores when human volunteers were replaced by a bot. There was a fig leaf placed over the naughty bits: a human still supervised the conversation, with the robot merely “advising” what to say. But down the line, we’re going to have a human “supervising” two, five, or ten parallel conversations — doing just enough to keep up the appearances while delivering faked compassion at a marvelous scale.
The same incentives will cause LLMs to swarm other types of online communications. Consider that many businesses don’t shy away from covert marketing: from paid product placement, to astroturfing on Reddit, to content-farmed blog articles, to outright fake reviews. Government agencies and political operatives regularly succumb to the same temptations when they feel the cause is just.
Still, despite the relative ubiquity of these strategies, they are kept in check by costs: past a modest threshold, you need to pay real humans to engage with others in believable ways. LLMs offer a solution, letting you effortlessly conjure millions of human-like personas that are broadly indistinguishable from real people and can seemingly live complex online lives for as long as needed — but by the end of the day, exist only to advance your hidden goal.
Marketing aside, the same toolkit would be indispensable for crime. Scams and spear phishing campaigns would reach new levels if one could perfectly tailor the communications to their marks’ professional and social backgrounds, and do so millions of times per day. Again, this is not science fiction: ChatGPT is already capable of “style transfer” that flawlessly adjusts the message to a person’s background. Its developers try to detect overtly malicious uses, but self-hosted implementations will be free of such constraints.
When it comes to the beneficial uses of LLMs, it is nearly a given that we will learn to depend on ChatGPT-style digital assistants to instantly retrieve, summarize, and apply the sum of human knowledge to any problem at hand. Some commentators raise concerns about accuracy, but I’m not buying that; the revolution is likely to happen even if Microsoft Clippy 2.0 occasionally makes a mistake or two.
But then, LLMs are successful specifically because of the culture of uninhibited and organic sharing on the internet. The emergence of LLM-based assistants threatens to sever this connection: in a world where Clippy 2.0 has all the answers, nobody will visit your website, ask you for advice, or send you a “thank you” note. Some creators might still find solace in helping the humanity in some abstract way, but most might give up — or flee to walled gardens where robots are not allowed to come.
A new breed of content licenses or other legal solutions to keep robots at bay might help preserve some degree of openness. The alternative is the internet of small, hermetic communities where members know each other, and the risk of drive-by robotic infiltration is easily kept in check.
I'm sure we'll know when LLMs replace human customer support: once we start seeing complete formal sentences with correct grammar :)
I really hope Neal Stephenson's "bogon" (from Anathem) catches on for this stuff.
On a more serious note, I think certificate chains (cleverly hidden behind an intuitive UI) that let people judge the source/path of information is the only hope.